Pseudo-RGB Data Augmentation for Improved Alzheimer's Disease Detection: An Analysis of CNN Performance

Alzheimer's disease, a progressive neurodegenerative disorder, presents substantial challenges in early diagnosis. It is characterized by cognitive impairment, damage to the memory, and tissue loss in the brain. Neuroimaging, particularly magnetic resonance imaging (MRI), has significantly contributed to comprehending brain functions and disorders over the past few decades. However, the limited availability of diverse and large-scale datasets due to privacy concerns impedes progress. Deep learning methodologies, especially convolutional neural networks (CNNs), have demonstrated considerable potential in analyzing neuroimaging data for diagnosing Alzheimer's disease (AD). Training a deep-learning model requires substantial data to yield precise outcomes and mitigate overfitting issues. This study addresses this limitation by evaluating the impact of advanced data augmentation techniques, a novel pseudo-RGB augmentation method, on AD classification using brain MRI scans. Five CNN architectures are used for performance analysis, including VGG16, ResNet50, MobileNet, InceptionV3, and Xception. We have applied the pseudo-coloring process to the dataset, and both standard and pseudo-RGB datasets are geometrically enhanced. Our findings reveal that the pseudo-RGB augmentation led to notable performance improvements across various metrics. On average, test accuracy improved by 2.81%, with the highest-performing model achieving 98.75\% accuracy on the pseudo-RGB dataset, compared to 97% on the standard dataset. The pseudo-RGB augmentation significantly reduced test loss by 26% and enhanced precision, recall, and F1 scores. Misclassification rates, particularly for the non-demented and mild-demented stages, also substantially improved. Our research shows the effectiveness of adding pseudo-coloring to achieve high accuracy in AD classification. Comparative evaluations with the existing literature highlight the usefulness and superiority of the proposed pseudo-RGB augmentation technique.